Late Gadolinium-Enhanced (LGE) Magnetic Resonance imaging (MRI) modality is of increasing relevance in clinical studies with atrial fibrillation patients due to its ability to detect scarring and fibrosis. However, segmentation of the atrial wall and cavities is a challenging task with respect to other heart chambers, especially on LGE-MRI images. The nnU-Net framework remains one of the top performing alternatives for medical image segmentation, whereas YOLOv8 provides fast identification of objects. In the context of the MBAS 2024 challenge, we employ a cascade of both models to segment the left and right atrial cavities and the biatrial myocardium. The spatial location of the biatrial region is identified using a YOLOv8 model and then the segmentation on the ROI is computed using a nnU-Net with residual connections. To provide robustness against magnetic field inhomogeneities, we add the logarithm of the LGE-MRI image as an extra input channel to the network, as a compromise between applying a time-consuming N4 bias correction algorithm and not correcting the images. Also, since there is a risk of the left and right myocardium merging in regions where they are very close together, the gold standard myocardium is split at training into left and right myocardium using an automatic method based on the Laplace equation (without any manual intervention). Both myocardial labels are converted into a single region at the output of the second network.